drivers/cpufreq/cpufreq_ondemand.c

   1 /*
   2  *  drivers/cpufreq/cpufreq_ondemand.c
   3  *
   4  *  Copyright (C)  2001 Russell King
   5  *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
   6  *                      Jun Nakajima <jun.nakajima@intel.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 as
  10  * published by the Free Software Foundation.
  11  */
  12
  13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15 #include <linux/cpufreq.h>
  16 #include <linux/init.h>
  17 #include <linux/kernel.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/kobject.h>
  20 #include <linux/module.h>
  21 #include <linux/mutex.h>
  22 #include <linux/percpu-defs.h>
  23 #include <linux/sysfs.h>
  24 #include <linux/tick.h>
  25 #include <linux/types.h>
  26
  27 #include "cpufreq_governor.h"
  28
  29 /* On-demand governor macors */
  30 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL         (10)
  31 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  32 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  33 #define MAX_SAMPLING_DOWN_FACTOR                (100000)
  34 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  35 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  36 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  37 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  38 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  39
  40 static struct dbs_data od_dbs_data;
  41 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
  42
  43 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  44 static struct cpufreq_governor cpufreq_gov_ondemand;
  45 #endif
  46
  47 static struct od_dbs_tuners od_tuners = {
  48         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  49         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
  50         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
  51         .ignore_nice = 0,
  52         .powersave_bias = 0,
  53 };
  54
  55 static void ondemand_powersave_bias_init_cpu(int cpu)
  56 {
  57         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
  58
  59         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
  60         dbs_info->freq_lo = 0;
  61 }
  62
  63 /*
  64  * Not all CPUs want IO time to be accounted as busy; this depends on how
  65  * efficient idling at a higher frequency/voltage is.
  66  * Pavel Machek says this is not so for various generations of AMD and old
  67  * Intel systems.
  68  * Mike Chan (androidlcom) calis this is also not true for ARM.
  69  * Because of this, whitelist specific known (series) of CPUs by default, and
  70  * leave all others up to the user.
  71  */
  72 static int should_io_be_busy(void)
  73 {
  74 #if defined(CONFIG_X86)
  75         /*
  76          * For Intel, Core 2 (model 15) andl later have an efficient idle.
  77          */
  78         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
  79                         boot_cpu_data.x86 == 6 &&
  80                         boot_cpu_data.x86_model >= 15)
  81                 return 1;
  82 #endif
  83         return 0;
  84 }
  85
  86 /*
  87  * Find right freq to be set now with powersave_bias on.
  88  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
  89  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
  90  */
  91 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
  92                 unsigned int freq_next, unsigned int relation)
  93 {
  94         unsigned int freq_req, freq_reduc, freq_avg;
  95         unsigned int freq_hi, freq_lo;
  96         unsigned int index = 0;
  97         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
  98         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
  99                                                    policy->cpu);
 100
 101         if (!dbs_info->freq_table) {
 102                 dbs_info->freq_lo = 0;
 103                 dbs_info->freq_lo_jiffies = 0;
 104                 return freq_next;
 105         }
 106
 107         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 108                         relation, &index);
 109         freq_req = dbs_info->freq_table[index].frequency;
 110         freq_reduc = freq_req * od_tuners.powersave_bias / 1000;
 111         freq_avg = freq_req - freq_reduc;
 112
 113         /* Find freq bounds for freq_avg in freq_table */
 114         index = 0;
 115         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 116                         CPUFREQ_RELATION_H, &index);
 117         freq_lo = dbs_info->freq_table[index].frequency;
 118         index = 0;
 119         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 120                         CPUFREQ_RELATION_L, &index);
 121         freq_hi = dbs_info->freq_table[index].frequency;
 122
 123         /* Find out how long we have to be in hi and lo freqs */
 124         if (freq_hi == freq_lo) {
 125                 dbs_info->freq_lo = 0;
 126                 dbs_info->freq_lo_jiffies = 0;
 127                 return freq_lo;
 128         }
 129         jiffies_total = usecs_to_jiffies(od_tuners.sampling_rate);
 130         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 131         jiffies_hi += ((freq_hi - freq_lo) / 2);
 132         jiffies_hi /= (freq_hi - freq_lo);
 133         jiffies_lo = jiffies_total - jiffies_hi;
 134         dbs_info->freq_lo = freq_lo;
 135         dbs_info->freq_lo_jiffies = jiffies_lo;
 136         dbs_info->freq_hi_jiffies = jiffies_hi;
 137         return freq_hi;
 138 }
 139
 140 static void ondemand_powersave_bias_init(void)
 141 {
 142         int i;
 143         for_each_online_cpu(i) {
 144                 ondemand_powersave_bias_init_cpu(i);
 145         }
 146 }
 147
 148 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 149 {
 150         if (od_tuners.powersave_bias)
 151                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 152         else if (p->cur == p->max)
 153                 return;
 154
 155         __cpufreq_driver_target(p, freq, od_tuners.powersave_bias ?
 156                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 157 }
 158
 159 /*
 160  * Every sampling_rate, we check, if current idle time is less than 20%
 161  * (default), then we try to increase frequency Every sampling_rate, we look for
 162  * a the lowest frequency which can sustain the load while keeping idle time
 163  * over 30%. If such a frequency exist, we try to decrease to this frequency.
 164  *
 165  * Any frequency increase takes it to the maximum frequency. Frequency reduction
 166  * happens at minimum steps of 5% (default) of current frequency
 167  */
 168 static void od_check_cpu(int cpu, unsigned int load_freq)
 169 {
 170         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 171         struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
 172
 173         dbs_info->freq_lo = 0;
 174
 175         /* Check for frequency increase */
 176         if (load_freq > od_tuners.up_threshold * policy->cur) {
 177                 /* If switching to max speed, apply sampling_down_factor */
 178                 if (policy->cur < policy->max)
 179                         dbs_info->rate_mult =
 180                                 od_tuners.sampling_down_factor;
 181                 dbs_freq_increase(policy, policy->max);
 182                 return;
 183         }
 184
 185         /* Check for frequency decrease */
 186         /* if we cannot reduce the frequency anymore, break out early */
 187         if (policy->cur == policy->min)
 188                 return;
 189
 190         /*
 191          * The optimal frequency is the frequency that is the lowest that can
 192          * support the current CPU usage without triggering the up policy. To be
 193          * safe, we focus 10 points under the threshold.
 194          */
 195         if (load_freq < (od_tuners.up_threshold - od_tuners.down_differential) *
 196                         policy->cur) {
 197                 unsigned int freq_next;
 198                 freq_next = load_freq / (od_tuners.up_threshold -
 199                                 od_tuners.down_differential);
 200
 201                 /* No longer fully busy, reset rate_mult */
 202                 dbs_info->rate_mult = 1;
 203
 204                 if (freq_next < policy->min)
 205                         freq_next = policy->min;
 206
 207                 if (!od_tuners.powersave_bias) {
 208                         __cpufreq_driver_target(policy, freq_next,
 209                                         CPUFREQ_RELATION_L);
 210                 } else {
 211                         int freq = powersave_bias_target(policy, freq_next,
 212                                         CPUFREQ_RELATION_L);
 213                         __cpufreq_driver_target(policy, freq,
 214                                         CPUFREQ_RELATION_L);
 215                 }
 216         }
 217 }
 218
 219 static void od_dbs_timer(struct work_struct *work)
 220 {
 221         struct od_cpu_dbs_info_s *dbs_info =
 222                 container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
 223         unsigned int cpu = dbs_info->cdbs.cpu;
 224         int delay, sample_type = dbs_info->sample_type;
 225
 226         mutex_lock(&dbs_info->cdbs.timer_mutex);
 227
 228         /* Common NORMAL_SAMPLE setup */
 229         dbs_info->sample_type = OD_NORMAL_SAMPLE;
 230         if (sample_type == OD_SUB_SAMPLE) {
 231                 delay = dbs_info->freq_lo_jiffies;
 232                 __cpufreq_driver_target(dbs_info->cdbs.cur_policy,
 233                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 234         } else {
 235                 dbs_check_cpu(&od_dbs_data, cpu);
 236                 if (dbs_info->freq_lo) {
 237                         /* Setup timer for SUB_SAMPLE */
 238                         dbs_info->sample_type = OD_SUB_SAMPLE;
 239                         delay = dbs_info->freq_hi_jiffies;
 240                 } else {
 241                         delay = delay_for_sampling_rate(od_tuners.sampling_rate
 242                                                 * dbs_info->rate_mult);
 243                 }
 244         }
 245
 246         schedule_delayed_work_on(cpu, &dbs_info->cdbs.work, delay);
 247         mutex_unlock(&dbs_info->cdbs.timer_mutex);
 248 }
 249
 250 /************************** sysfs interface ************************/
 251
 252 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 253                                       struct attribute *attr, char *buf)
 254 {
 255         return sprintf(buf, "%u\n", od_dbs_data.min_sampling_rate);
 256 }
 257
 258 /**
 259  * update_sampling_rate - update sampling rate effective immediately if needed.
 260  * @new_rate: new sampling rate
 261  *
 262  * If new rate is smaller than the old, simply updaing
 263  * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
 264  * original sampling_rate was 1 second and the requested new sampling rate is 10
 265  * ms because the user needs immediate reaction from ondemand governor, but not
 266  * sure if higher frequency will be required or not, then, the governor may
 267  * change the sampling rate too late; up to 1 second later. Thus, if we are
 268  * reducing the sampling rate, we need to make the new value effective
 269  * immediately.
 270  */
 271 static void update_sampling_rate(unsigned int new_rate)
 272 {
 273         int cpu;
 274
 275         od_tuners.sampling_rate = new_rate = max(new_rate,
 276                         od_dbs_data.min_sampling_rate);
 277
 278         for_each_online_cpu(cpu) {
 279                 struct cpufreq_policy *policy;
 280                 struct od_cpu_dbs_info_s *dbs_info;
 281                 unsigned long next_sampling, appointed_at;
 282
 283                 policy = cpufreq_cpu_get(cpu);
 284                 if (!policy)
 285                         continue;
 286                 if (policy->governor != &cpufreq_gov_ondemand) {
 287                         cpufreq_cpu_put(policy);
 288                         continue;
 289                 }
 290                 dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 291                 cpufreq_cpu_put(policy);
 292
 293                 mutex_lock(&dbs_info->cdbs.timer_mutex);
 294
 295                 if (!delayed_work_pending(&dbs_info->cdbs.work)) {
 296                         mutex_unlock(&dbs_info->cdbs.timer_mutex);
 297                         continue;
 298                 }
 299
 300                 next_sampling = jiffies + usecs_to_jiffies(new_rate);
 301                 appointed_at = dbs_info->cdbs.work.timer.expires;
 302
 303                 if (time_before(next_sampling, appointed_at)) {
 304
 305                         mutex_unlock(&dbs_info->cdbs.timer_mutex);
 306                         cancel_delayed_work_sync(&dbs_info->cdbs.work);
 307                         mutex_lock(&dbs_info->cdbs.timer_mutex);
 308
 309                         schedule_delayed_work_on(dbs_info->cdbs.cpu,
 310                                         &dbs_info->cdbs.work,
 311                                         usecs_to_jiffies(new_rate));
 312
 313                 }
 314                 mutex_unlock(&dbs_info->cdbs.timer_mutex);
 315         }
 316 }
 317
 318 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 319                                    const char *buf, size_t count)
 320 {
 321         unsigned int input;
 322         int ret;
 323         ret = sscanf(buf, "%u", &input);
 324         if (ret != 1)
 325                 return -EINVAL;
 326         update_sampling_rate(input);
 327         return count;
 328 }
 329
 330 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 331                                    const char *buf, size_t count)
 332 {
 333         unsigned int input;
 334         int ret;
 335
 336         ret = sscanf(buf, "%u", &input);
 337         if (ret != 1)
 338                 return -EINVAL;
 339         od_tuners.io_is_busy = !!input;
 340         return count;
 341 }
 342
 343 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 344                                   const char *buf, size_t count)
 345 {
 346         unsigned int input;
 347         int ret;
 348         ret = sscanf(buf, "%u", &input);
 349
 350         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 351                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 352                 return -EINVAL;
 353         }
 354         od_tuners.up_threshold = input;
 355         return count;
 356 }
 357
 358 static ssize_t store_sampling_down_factor(struct kobject *a,
 359                         struct attribute *b, const char *buf, size_t count)
 360 {
 361         unsigned int input, j;
 362         int ret;
 363         ret = sscanf(buf, "%u", &input);
 364
 365         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 366                 return -EINVAL;
 367         od_tuners.sampling_down_factor = input;
 368
 369         /* Reset down sampling multiplier in case it was active */
 370         for_each_online_cpu(j) {
 371                 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 372                                 j);
 373                 dbs_info->rate_mult = 1;
 374         }
 375         return count;
 376 }
 377
 378 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 379                                       const char *buf, size_t count)
 380 {
 381         unsigned int input;
 382         int ret;
 383
 384         unsigned int j;
 385
 386         ret = sscanf(buf, "%u", &input);
 387         if (ret != 1)
 388                 return -EINVAL;
 389
 390         if (input > 1)
 391                 input = 1;
 392
 393         if (input == od_tuners.ignore_nice) { /* nothing to do */
 394                 return count;
 395         }
 396         od_tuners.ignore_nice = input;
 397
 398         /* we need to re-evaluate prev_cpu_idle */
 399         for_each_online_cpu(j) {
 400                 struct od_cpu_dbs_info_s *dbs_info;
 401                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 402                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 403                                                 &dbs_info->cdbs.prev_cpu_wall);
 404                 if (od_tuners.ignore_nice)
 405                         dbs_info->cdbs.prev_cpu_nice =
 406                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 407
 408         }
 409         return count;
 410 }
 411
 412 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 413                                     const char *buf, size_t count)
 414 {
 415         unsigned int input;
 416         int ret;
 417         ret = sscanf(buf, "%u", &input);
 418
 419         if (ret != 1)
 420                 return -EINVAL;
 421
 422         if (input > 1000)
 423                 input = 1000;
 424
 425         od_tuners.powersave_bias = input;
 426         ondemand_powersave_bias_init();
 427         return count;
 428 }
 429
 430 show_one(od, sampling_rate, sampling_rate);
 431 show_one(od, io_is_busy, io_is_busy);
 432 show_one(od, up_threshold, up_threshold);
 433 show_one(od, sampling_down_factor, sampling_down_factor);
 434 show_one(od, ignore_nice_load, ignore_nice);
 435 show_one(od, powersave_bias, powersave_bias);
 436
 437 define_one_global_rw(sampling_rate);
 438 define_one_global_rw(io_is_busy);
 439 define_one_global_rw(up_threshold);
 440 define_one_global_rw(sampling_down_factor);
 441 define_one_global_rw(ignore_nice_load);
 442 define_one_global_rw(powersave_bias);
 443 define_one_global_ro(sampling_rate_min);
 444
 445 static struct attribute *dbs_attributes[] = {
 446         &sampling_rate_min.attr,
 447         &sampling_rate.attr,
 448         &up_threshold.attr,
 449         &sampling_down_factor.attr,
 450         &ignore_nice_load.attr,
 451         &powersave_bias.attr,
 452         &io_is_busy.attr,
 453         NULL
 454 };
 455
 456 static struct attribute_group od_attr_group = {
 457         .attrs = dbs_attributes,
 458         .name = "ondemand",
 459 };
 460
 461 /************************** sysfs end ************************/
 462
 463 define_get_cpu_dbs_routines(od_cpu_dbs_info);
 464
 465 static struct od_ops od_ops = {
 466         .io_busy = should_io_be_busy,
 467         .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 468         .powersave_bias_target = powersave_bias_target,
 469         .freq_increase = dbs_freq_increase,
 470 };
 471
 472 static struct dbs_data od_dbs_data = {
 473         .governor = GOV_ONDEMAND,
 474         .attr_group = &od_attr_group,
 475         .tuners = &od_tuners,
 476         .get_cpu_cdbs = get_cpu_cdbs,
 477         .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 478         .gov_dbs_timer = od_dbs_timer,
 479         .gov_check_cpu = od_check_cpu,
 480         .gov_ops = &od_ops,
 481 };
 482
 483 static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
 484                 unsigned int event)
 485 {
 486         return cpufreq_governor_dbs(&od_dbs_data, policy, event);
 487 }
 488
 489 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 490 static
 491 #endif
 492 struct cpufreq_governor cpufreq_gov_ondemand = {
 493         .name                   = "ondemand",
 494         .governor               = od_cpufreq_governor_dbs,
 495         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 496         .owner                  = THIS_MODULE,
 497 };
 498
 499 static int __init cpufreq_gov_dbs_init(void)
 500 {
 501         u64 idle_time;
 502         int cpu = get_cpu();
 503
 504         mutex_init(&od_dbs_data.mutex);
 505         idle_time = get_cpu_idle_time_us(cpu, NULL);
 506         put_cpu();
 507         if (idle_time != -1ULL) {
 508                 /* Idle micro accounting is supported. Use finer thresholds */
 509                 od_tuners.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 510                 od_tuners.down_differential = MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 511                 /*
 512                  * In nohz/micro accounting case we set the minimum frequency
 513                  * not depending on HZ, but fixed (very low). The deferred
 514                  * timer might skip some samples if idle/sleeping as needed.
 515                 */
 516                 od_dbs_data.min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 517         } else {
 518                 /* For correct statistics, we need 10 ticks for each measure */
 519                 od_dbs_data.min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 520                         jiffies_to_usecs(10);
 521         }
 522
 523         return cpufreq_register_governor(&cpufreq_gov_ondemand);
 524 }
 525
 526 static void __exit cpufreq_gov_dbs_exit(void)
 527 {
 528         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 529 }
 530
 531 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 532 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 533 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 534         "Low Latency Frequency Transition capable processors");
 535 MODULE_LICENSE("GPL");
 536
 537 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 538 fs_initcall(cpufreq_gov_dbs_init);
 539 #else
 540 module_init(cpufreq_gov_dbs_init);
 541 #endif
 542 module_exit(cpufreq_gov_dbs_exit);